Research On Quantitative Investment Strategy Model: Based On Deep Neural Network

The rapid development of computers has made quantitative investment strategy an important investment tool.Renaissance Technologies,the best known quantitative investment management funds founded by James Simons,created an impressive 35% annual average profit from 1989 to 2009 earning this approach increasing approval from investors.At the same time,with the advent of deep neural network(DNN)in 2006,deep learning has made impressing achievements in the field of image recognition,speech recognition,text-based emotion detection and recognition etc.Later on,scholars started research on applying deep learning system to quantitate investing and tried to use artificial intelligence to interpret the nature of financial markets.Applying deep learning to quantitative invest ing is a relatively new direction.In this study,through reviewing and summarizing related theories,this paper first selected 30 companies that have been long listed in A-share market,are at good liquidity level,and have certain representativeness.The author chose and normalized 60 factors including price,indicators as well as market factor,a noble factor.Then,the author employed Principal Component Analysis(PCA),Auto Encoder,and Restricted Boltzmann Machine(RBM)for feature extraction and used Logistic Regression model to compare the prediction effect of these feature data.After that the author compared the selected model and feature data with the original data,and input them in DNN model.The optimal model was selected by comparing the three objective functions.The the author constructed a Deep Neural Network-Logistic Regression(DNN-LR)Model to compare Artificial Neural Network and Deep Neural Network Model and found the former moedl has higher accuracy.Finally,use the model from the previous step to make predictions to stock price movements.Construct a quantitative investment strategy with the predictions and compare it with buy and hold strategy,classic index and CSI 300 index.On the basis of those strategies,the author further built a portfolio to calculate profitability,maximum retracement,Sharpe ratio and other indicators and returns under different transaction costs.The experimental results show following aspects: first,PCA is the best model among feature extraction methods.However,compared with the original data,the prediction results from original data are still the best showing that feature extraction methods can't well extract features of financial data;Second,among DNN learning approaches,the best model structure is the objective function with the normalized mean square error.The average prediction accuracy of this model reached 61.40%;Third,the quantitative investment strategy built on the above basis,whether the market is in rising,falling,fluctuating trends,the overall market performance far exceeds buy and hold strategy and classic index strategy.The portfolio based on the quantitative strategy has even greater performance.Finally,it is calculated that the strategy is relatively insensitive to transaction costs,and thus can withstand higher costs.All above shows that the quantitative strategy of this article has a certain practical value.The innovation of this paper lies in the construction of a deep neural network-logistic regression model(DNN-LR),which combines the features of the two models and improves the prediction accuracy of stock prices.At the same time,this article set a method of dynamic balance adjustment position backtesting and found that the method's return rate is higher than the single stock's strategy rate of return.